Exploring the tradeoffs between programmability and efficiency in data-parallel accelerators